PaperlayerView exact revision on arXiv

Source-linked AI summary

Principles alone cannot guarantee ethical AI

Brent Mittelstadt

arXiv:1906.06668v2cs.CYcs.AI

TL;DR

High-level AI ethics principles may conceal unresolved political and normative disagreements, with limited evidence that principlism will translate effectively from medicine to AI. The paper reviews AI Ethics initiatives and compares their approach with principlism’s implementation and impact in medicine, finding four features that constrain its likely impact on AI design and governance.

  • Problem

    AI Ethics initiatives largely offer vague principles without specific recommendations or resolution of fundamental normative and political tensions.

  • Method

    The paper reviews AI Ethics initiatives and uses research on medical principlism to critically assess principled AI ethics.

  • Results

    AI development lacks common aims and fiduciary duties, professional history and norms, proven translation methods, and robust accountability mechanisms.

  • Takeaways & Limitations

    Consensus on shared principles is insufficient to guarantee trustworthy or ethical AI, making resolution of contested concepts the central work ahead.

  • Takeaways & Limitations

    Principlism has not been an unqualified success in medicine, so its influence there does not establish comparable success for AI.

Abstract

from arXiv · show

AI Ethics is now a global topic of discussion in academic and policy circles. At least 84 public-private initiatives have produced statements describing high-level principles, values, and other tenets to guide the ethical development, deployment, and governance of AI. According to recent meta-analyses, AI Ethics has seemingly converged on a set of principles that closely resemble the four classic principles of medical ethics. Despite the initial credibility granted to a principled approach to AI Ethics by the connection to principles in medical ethics, there are reasons to be concerned about its future impact on AI development and governance. Significant differences exist between medicine and AI development that suggest a principled approach in the latter may not enjoy success comparable to the former. Compared to medicine, AI development lacks (1) common aims and fiduciary duties, (2) professional history and norms, (3) proven methods to translate principles into practice, and (4) robust legal and professional accountability mechanisms. These differences suggest we should not yet celebrate consensus around high-level principles that hide deep political and normative disagreement.

1 Introduction

At least 84 initiatives have articulated high-level AI ethics principles, many resembling medical ethics, but their vague guidance and AI development’s differences from medicine raise doubts about principlism’s practical impact. The paper assesses these initiatives’ strategies for embedding ethics in AI development and governance using lessons from medical principlism.

  • 1 Introduction: At least 84 AI Ethics initiatives have published high-level principles, values, or abstract requirements, often envisioning translation into design requirements, governance frameworks, and professional codes.These initiatives also aim to focus debate and raise awareness of AI’s ethical challenges.
  • 1 Introduction: Critics argue that industry-sponsored initiatives can function as virtue signalling, while their vague high-level statements provide few specific action-guiding requirements in practice.The initiatives may thereby delay regulation or pre-emptively direct debate toward abstract problems and technical solutions.
  • 1 Introduction: Many AI Ethics initiatives have converged on principles closely resembling medicine’s four classic principles, including autonomy, harm prevention, fairness, and explicability.The convergence was endorsed by the OECD and the European Commission’s HLEG.
  • 1 Introduction: Important differences between medicine and AI development suggest that a principled approach to AI ethics may not achieve medicine’s success despite the comparison’s initial credibility.The paper frames these differences as reasons for concern about AI Ethics’ future impact.
  • 1 Introduction: The paper critically assesses current initiatives’ recommendations for embedding ethics in AI development and governance by applying prior work on principlism’s implementation and impact in medicine.It reviews existing initiatives’ outputs to identify their proposed embedding strategies.

2 The challenges of a principled approach to AI Ethics

A principled approach to AI Ethics may have limited impact because AI development lacks medicine’s shared aims, professional norms, implementation methods, and accountability mechanisms. High-level consensus can obscure political disagreement and offers little justification for context-specific requirements or unified implementation.

  • 2.1 Common aims and fiduciary duties: AI development lacks medicine’s common aims and fiduciary duties, leaving developers without equivalent obligations to prioritize affected individuals over organizational interests.Developers face pressures to reduce costs, increase profits, and prioritize company interests, while AI lacks an equivalent patient whose interests receive initial primacy.
  • 2.2 Professional history and norms: AI development lacks medicine’s professional history and norms, while its multidisciplinary, multinational teams make shared standards difficult to establish and apply.Medical codes developed over time into detailed standards covering conduct, behavior, and diverse practices; AI’s distributed development makes effects less observable to developers.
  • 2.3 Translating principles into practice: 84 public-private initiatives have produced high-level AI Ethics statements, but abstract principles such as fairness and dignity can conceal genuinely conflicting political and ethical interpretations.These contested concepts may generate different practical requirements, with ambiguity potentially enabling contextual specification or masking fundamental disagreement.
  • 2.3 Translating principles into practice: AI development lacks empirically proven methods for translating principles into mid-level norms and low-level requirements, so consensus on principles does not transfer to specific contexts.Developers must specify contested concepts without a clear roadmap, and practical requirements must also be embedded in design and development processes.
  • 2.3 Translating principles into practice: Commercial AI development faces added obstacles to ethical implementation because stakeholder participation, embedded ethicists, and conflict resolution impose work and costs that may be discarded for commercial reasons.AI is often developed behind closed doors without public representation, and ethical considerations may be dropped when they conflict with commercial priorities.
  • 2.4 Legal and professional accountability: AI Ethics has relatively weak legal and professional accountability mechanisms compared with medicine, and serious long-term commitment to self-regulation cannot be assumed.Medicine uses malpractice law, licensing, certification, ethics committees, and professional boards to uphold standards and provide redress.

3 Where should AI Ethics go from here?

The shortcomings of AI development compared with medicine create significant challenges for implementing AI Ethics. Consensus on high-level principles should therefore not be treated as sufficient to guarantee trustworthy or ethical AI.

  • 3 Where should AI Ethics go from here?: AI development’s comparative shortcomings to medicine raise significant challenges for implementing AI Ethics.The paper identifies differences involving aims and duties, professional norms, translation into practice, and accountability mechanisms.
  • 3 Where should AI Ethics go from here?: Self-regulatory codes without clearly defined and enforceable obligations can benefit developers through trustworthiness and reputation without imposing costs.The passage argues that signing up to such codes costs developers nothing while producing immediate reputational benefits.
  • 3 Where should AI Ethics go from here?: Consensus around high-level principles should not be celebrated because it can hide deep political and normative disagreement.Shared principles alone are not enough to guarantee trustworthy or ethical AI.
  • 3 Where should AI Ethics go from here?: Without a fundamental regulatory shift, translating principles into practice will remain competitive rather than cooperative.The passage describes principles as vacuous until they are tested, when their true costs and value become apparent.

1. Clearly define sustainable pathways to impact

Principled AI initiatives need clearly defined long-term aims and pathways to impact, supported by cooperative oversight, visible accountability, and clear implementation and review processes.

  • Principled approaches require cooperative oversight to keep translated norms and requirements fit for purpose and impactful over time.
  • Long-term aims and pathways to impact should be clarified, alongside binding accountability structures and clear sectoral and organisational implementation and review processes.

2. Support ‘bottom-up’ AI Ethics in the private sector

Because AI encompasses diverse technologies, top-down ethics approaches are especially difficult; bottom-up case studies of production systems can help develop principles and professional standards.

  • Support ‘bottom-up’ AI Ethics in the private sector: AI’s technological diversity makes top-down ethics approaches uniquely difficult, requiring complementary bottom-up case studies of production systems.These case studies address the limitations of generalist approaches in a diverse field.
  • Support ‘bottom-up’ AI Ethics in the private sector: Collaborative assessment of local practices can specify principles and establish precedents that advance professional standards.Local practices provide a basis for developing more specific guidance.
  • Support ‘bottom-up’ AI Ethics in the private sector: Novel cases expose new AI Ethics challenges and can move the field beyond well-worn examples.The passage characterizes these new challenges as necessary for advancing the field.

3. License developers of high-risk AI

The section proposes formally establishing AI development as a profession with standing equivalent to other high-risk professions. Initial licensing efforts could focus on developers of elevated-risk or public-sector systems, including facial recognition for policing.

  • 3. License developers of high-risk AI: AI development may need formal recognition as a profession equivalent in standing to other high-risk professions.This is presented as a way to encourage long-term recognition of ethical commitments.
  • 3. License developers of high-risk AI: It is a regulatory oddity that public-service professions are licensed while developers of technical systems to augment or replace human activities are not.
  • 3. License developers of high-risk AI: Initial licensing initiatives could target developers of elevated-risk systems or systems built for the public sector.One example is facial recognition designed for policing.

4. Shift from professional ethics to organisational ethics

Many AI Ethics initiatives resemble professional codes focused on individual designers’ requirements, behaviours, and values. This framing leaves applications and organisational interests largely unquestioned while shifting attention from unethical organisations and business models to individual transgressions.

  • 4. Shift from professional ethics to organisational ethics: AI Ethics initiatives often resemble professional codes addressing design requirements, behaviours, and values of individual professions.The passage contrasts this individual-profession focus with scrutiny of broader organisational and business interests.
  • 4. Shift from professional ethics to organisational ethics: The legitimacy of particular applications and their underlying business and organisational interests remains largely unquestioned.The passage identifies this lack of scrutiny as a central limitation of the professional-ethics framing.
  • 4. Shift from professional ethics to organisational ethics: This approach redirects debate toward unethical individuals’ transgressions and away from collective organisational failures and unethical business models.The passage presents this redirection as a convenient consequence of focusing on professional ethics.

5. Pursue ethics as a process, not technological solutionism

AI Ethics should be pursued as an ongoing process rather than reduced to technical fixes for ethical challenges. Because AI is embedded in practices’ political and ethical dimensions, principled disagreement should be expected while principles are translated into practice.

  • Technical solutionism: Technical and design expertise often frames ethical challenges such as privacy and fairness as addressable through technical fixes, but technical definitions and explanations are rarely proposed.The IEEE’s ‘Ethically-Aligned Design’ initiative is an exception, though its commercial impact and uptake remain uncertain.
  • Technical solutionism: Framing ethical challenges as design flaws keeps them fundamentally technical and shields them from democratic intervention.AI’s capacity to replace or augment human expertise entangles it with the ethical and political dimensions of the practices in which it is embedded.
  • Ethics as a process: Ethics is a process, not a destination, so principled disagreements should be expected and welcomed rather than treated as failures requiring resolution.The work of AI Ethics is to translate and implement lofty principles while beginning to understand AI’s real ethical challenges.

Competing Interests

The author reports receiving reimbursement for conference-related travel from funding provided by DeepMind Technologies Limited.

  • Competing Interests: The author previously received reimbursement for conference-related travel funded by DeepMind Technologies Limited.

Tables

The tables present characteristics of a formal profession and eight questions for assessing the value and practical implications of AI Ethics principles.

  • Table 2 asks who wrote an AI Ethics statement, how it was written, whom it targets, and what purpose it serves.
  • It asks why the statement should be followed, how it should be implemented, and how conflicting interpretations of contested concepts should be resolved.
  • It asks how adherence will be assessed, what happens when the statement is not followed, and how disagreements or clarification questions can be raised.
Loading 1906.06668v2…